Rakshamedra/FLAIR-HUB

---
license: etalab-2.0
size_categories:
- 100K<n<1M
task_categories:
- image-segmentation
pretty_name: FLAIR-HUB
tags:
- Multimodal
- Earth Observation
- Remote Sensing
- Aerial
- Satellite
- Environement
- LandCover
- Agriculture
---

# FLAIR-HUB : Large-scale Multimodal Dataset for Land Cover and Crop Mapping

FLAIR-HUB builds upon and includes the FLAIR#1 and FLAIR#2 datasets, expanding them into a unified, large-scale, multi-sensor land-cover resource with very-high-resolution
annotations. Spanning over 2,500 km² of diverse French ecoclimates and landscapes, it features 63 billion hand-annotated pixels across 19 land-cover and
23 crop type classes.<br>
The dataset integrates complementary data sources including aerial imagery, SPOT and Sentinel satellites, surface models, and historical aerial photos,
offering rich spatial, spectral, and temporal diversity. FLAIR-HUB supports the development of semantic segmentation, multimodal fusion, and self-supervised
learning methods, and will continue to grow with new modalities and annotations.

<p align="center"><img src="datacard_imgs/FLAIR-HUB_Overview.png" alt="" style="width:70%;max-width:1600px;" /></p>
<hr>

## 🔗 Links

📄 <a href="https://arxiv.org/abs/2506.07080" target="_blank"><b>Dataset Preprint</b></a><br>
📄 <a href="https://huggingface.co/papers/2508.10894" target="_blank"><b>MAESTRO Paper (using this dataset)</b></a><br>
📁 <a href="https://storage.gra.cloud.ovh.net/v1/AUTH_366279ce616242ebb14161b7991a8461/defi-ia/flair_hub/FLAIR-HUB_TOY_DATASET.zip" target="_blank"><b>Toy dataset (~750MB) -direct download-</b></a><br>
💻 <a href="https://github.com/IGNF/FLAIR-HUB" target="_blank"><b>Source Code (GitHub)</b></a><br>
💻 <a href="https://github.com/ignf/maestro" target="_blank"><b>MAESTRO Code (GitHub, uses this dataset)</b></a><br>
🏠 <a href="https://ignf.github.io/FLAIR/" target="_blank"><b>FLAIR datasets page </b></a><br>
✉️ <a href="mailto:[e-mail]"><b>Contact Us</b></a> – [e-mail] – Questions or collaboration inquiries welcome!<br>
<hr>

## 🎯 Key Figures

<table>
<tr><td>🗺️</td><td>ROI / Area Covered</td><td>➡️ 2,822 ROIs / 2,528 km²</td></tr>
<tr><td>🧠</td><td>Modalities</td><td>➡️ 6 modalities</td></tr>
<tr><td>🏛️</td><td>Departments (France)</td><td>➡️ 74</td></tr>
<tr><td>🧩</td><td>AI Patches (512×512 px @ 0.2m)</td><td>➡️ 241,100</td></tr>
<tr><td>🖼️</td><td>Annotated Pixels</td><td>➡️ 63.2 billion</td></tr>
<tr><td>🛰️</td><td>Sentinel-2 Acquisitions</td><td>➡️ 256,221</td></tr>
<tr><td>📡</td><td>Sentinel-1 Acquisitions</td><td>➡️ 532,696</td></tr>
<tr><td>📁</td><td>Total Files</td><td>➡️ ~2.5 million</td></tr>
<tr><td>💾</td><td>Total Dataset Size</td><td>➡️ ~750 GB</td></tr>
</table>
<hr>

## 🗃️ Dataset Structure

```
data/
├── DOMAIN_SENSOR_DATATYPE/
│ ├── ROI/
│ │ ├── <Patch>.tif # image file
│ │ ├── <Patch>.tif
| | ├── ...
│ └── ...
├── ...
├── DOMAIN_SENSOR_LABEL-XX/
│ ├── ROI/
│ │ ├── <Patch>.tif # supervision file
│ │ ├── <Patch>.tif
│ └── ...
├── ...
└── GLOBAL_ALL_MTD/
├── GLOABAL_SENSOR_MTD.gpkg # metadata file
├── GLOABAL_SENSOR_MTD.gpkg
└── ...
```

## 🗂️ Data Modalities Overview

<center>
<table>
<thead>
<tr>
<th>Modality</th>
<th>Description</th>
<th style="text-align: center">Resolution / Format</th>
<th style="text-align: center">Metadata</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>BD ORTHO (AERIAL_RGBI)</strong></td>
<td>Orthorectified aerial images with 4 bands (R, G, B, NIR).</td>
<td style="text-align: center">20 cm, 8-bit unsigned</td>
<td style="text-align: center">Radiometric stats, acquisition dates/cameras</td>
</tr>
<tr>
<td><strong>BD ORTHO HISTORIQUE (AERIAL-RLT_PAN)</strong></td>
<td>Historical panchromatic aerial images (1947–1965), resampled.</td>
<td style="text-align: center">~40 cm, real: 0.4–1.2 m, 8-bit</td>
<td style="text-align: center">Dates, original image references</td>
</tr>
<tr>
<td><strong>ELEVATION (DEM_ELEV)</strong></td>
<td>Elevation data with DSM (surface) and DTM (terrain) channels.</td>
<td style="text-align: center">DSM: 20 cm, DTM: 1 m, Float32</td>
<td style="text-align: center">Object heights via DSM–DTM difference</td>
</tr>
<tr>
<td><strong>SPOT (SPOT_RGBI)</strong></td>
<td>SPOT 6-7 satellite images, 4 bands, calibrated reflectance.</td>
<td style="text-align: center">1.6 m (resampled)</td>
<td style="text-align: center">Acquisition dates, radiometric stats</td>
</tr>
<tr>
<td><strong>SENTINEL-2 (SENTINEL2_TS)</strong></td>
<td>Annual time series with 10 spectral bands, calibrated reflectance.</td>
<td style="text-align: center">10.24 m (resampled)</td>
<td style="text-align: center">Dates, radiometric stats, cloud/snow masks</td>
</tr>
<tr>
<td><strong>SENTINEL-1 ASC/DESC (SENTINEL1-XXX_TS)</strong></td>
<td>Radar time series (VV, VH), SAR backscatter (σ0).</td>
<td style="text-align: center">10.24 m (resampled)</td>
<td style="text-align: center">Stats per ascending/descending series</td>
</tr>
<tr>
<td><strong>LABELS CoSIA (AERIAL_LABEL-COSIA)</strong></td>
<td>Land cover labels from aerial RGBI photo-interpretation.</td>
<td style="text-align: center">20 cm, 15–19 classes</td>
<td style="text-align: center">Aligned with BD ORTHO, patch statistics</td>
</tr>
<tr>
<td><strong>LABELS LPIS (ALL_LABEL-LPIS)</strong></td>
<td>Crop type data from CAP declarations, hierarchical class structure.</td>
<td style="text-align: center">20 cm</td>
<td style="text-align: center">Aligned with BD ORTHO, may differ from CoSIA</td>
</tr>
</tbody>
</table>
</center>

<p align="center"><img src="datacard_imgs/FLAIR-HUB_Patches_Hori.png" alt="" style="width:100%;max-width:1300px;" /></p>
<hr>

## 🏷️ Supervision

FLAIR-HUB includes two complementary supervision sources: AERIAL_LABEL-COSIA, a high-resolution land cover annotation derived from expert photo-interpretation
of RGBI imagery, offering pixel-level precision across 19 classes; and AERIAL_LABEL-LPIS, a crop-type annotation based on farmer-declared parcels
from the European Common Agricultural Policy, structured into a three-level taxonomy of up to 46 crop classes. While COSIA reflects actual land cover,
LPIS captures declared land use, and the two differ in purpose, precision, and spatial alignment.

<p align="center"><img src="datacard_imgs/FLAIR-HUB_Labels.png" alt="" style="width:70%;max-width:1300px;" /></p>
<hr>

## 🌍 Spatial partition

FLAIR-HUB uses an <b>official split for benchmarking, corresponding to the split_1 fold</b>.

</div>

<div style="flex: 60%; margin: auto;"">
<table border="1">
<tr>
<th><font color="#c7254e">TRAIN / VALIDATION </font></th>
<td>D004, D005, D006, D007, D008, D009, D010, D011, D013, D014, D016, D017, D018, D020, D021, D023, D024047, D025039, D029,
D030, D031, D032, D033, D034, D035, D037, D038, D040, D041, D044, D045, D046, D049, D051, D052, D054057, D055, D056, D058, D059062,
D060, D063, D065, D066, D067, D070, D072, D074, D077, D078, D080, D081, D086, D091</td>
</tr>
<tr>
<th><font color="#c7254e">TEST</font></th>
<td>D012, D015, D022, D026, D036, D061, D064, D068, D069, D071, D073, D075, D076, D083, D084, D085</td>
</tr>
</table>
</div>
</div>

<p align="center"><img src="datacard_imgs/FLAIR-HUB_splits_oneline.png" alt="" style="width:80%;max-width:1300px;" /></p>
<hr>

## 🏆 Bechmark scores

Several model configurations were trained (see the accompanying data paper).
The best-performing configurations for both land-cover and crop-type classification tasks are summarized below:

<div align="center">

Task | Model ID | mIoU | O.A.
:------------ | :------------- | :-----------| :---------
🗺️ Land-cover | LC-L | 65.8 | 78.2
🌾 Crop-types | LPIS-I | 39.2 | 87.2

</div>

The **Model ID** can be used to retrieve the corresponding pre-trained model from the FLAIR-HUB-MODELS collection.

🗺️ Land-cover

| Model ID | Aerial VHR | Elevation | SPOT | S2 t.s. | S1 t.s. | Historical | PARA. | EP. | O.A. | mIoU |
|----------|------------|-----------|------|---------|---------|------------|--------|-----|------|------|
| LC-A | ✓ | | | | | | 89.4 | 79 | 77.5 | 64.1 |
| LC-B | ✓ | ✓ | | | | | 181.4 | 124 | 78.1 | 65.1 |
| LC-C | ✓ | ✓ | ✓ | | | | 270.6 | 129 | 78.2 | 65.2 |
| LC-D | ✓ | | | ✓ | | | 93.9 | 85 | 77.6 | 64.7 |
| LC-E | ✓ | | | | ✓ | | 95.8 | 98 | 77.7 | 64.5 |
| LC-F | ✓ | | | ✓ | ✓ | | 97.7 | 64 | 77.7 | 64.9 |
| LC-G | | | | ✓ | | | 0.9 | 89 | 57.8 | 34.2 |
| LC-H | | | | | ✓ | | 1.8 | 106 | 54.5 | 28.2 |
| LC-I | | | ✓ | | | | 89.2 | 94 | 64.1 | 43.5 |
| LC-J | | ✓ | | | | | 89.4 | 97 | 67.4 | 51.2 |
| LC-K | ✓ | | | | | ✓ | 181.4 | 45 | 77.6 | 64.3 |
| LC-L | ✓ | ✓ | ✓ | ✓ | ✓ | | 276.4 | 121 | **78.2** | **65.8** |
| LC-ALL | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | 365.8 | 129 | **78.2** | 65.6 |

🌾 Crop-types

| Model ID | Aerial VHR | SPOT | S2 t.s. | S1 t.s. | PARA. | EP. | O.A. | mIoU |
|----------|------------|------|---------|---------|--------|-----|------|------|
| **LV.1 - 23 classes (2 classes removed)** |||||||||
| LPIS-A | ✓ | | | | 89.4 | 91 | 86.6 | 24.4 |
| LPIS-B | ✓ | ✓ | | | 181.2 | 99 | 87.1 | 26.1 |
| LPIS-C | ✓ | | ✓ | | 93.9 | 100 | 87.5 | 29.8 |
| LPIS-D | ✓ | | ✓ | ✓ | 97.7 | 45 | **88.0** | 36.1 |
| LPIS-E | ✓ | ✓ | ✓ | | 183.1 | 46 | 87.6 | 30.3 |
| LPIS-F | | | ✓ | | 0.9 | 61 | 85.3 | 23.8 |
| LPIS-G | | | | ✓ | 1.8 | 77 | 84.5 | 18.1 |
| LPIS-H | | | ✓ | ✓ | 2.8 | 61 | 84.9 | 23.8 |
| LPIS-I | | ✓ | ✓ | ✓ | 97.5 | 49 | 87.2 | **39.2** |
| LPIS-J | ✓ | ✓ | ✓ | ✓ | 186.9 | 53 | **88.0** | 35.4 |
| LPIS-K | | ✓ | | | 89.2 | 14 | 84.5 | 15.1 |

<hr>

## 🔎 Filter dataset with the FLAIR-HUB Dataset Browser

A small desktop GUI to browse and download subsets
of the **IGNF/FLAIR-HUB** dataset from Hugging Face with filters for: Domain, Year, Modality or Data type.

Requirements:

- Python **3.9+**
- Tkinter (usually included; on Linux you may need: sudo apt-get install python3-tk)
- Python packages: pip install `huggingface_hub`

Run:
1. Download the file `flair-hub-HF-dl.py` from the *Files* section of this dataset.
2. In a terminal: ```pip install huggingface_hub```
3. Launch: ```python flair-hub-HF-dl.py```

<hr>

## ✨ MAESTRO basecode

This dataset is extensively used by the [MAESTRO model](https://huggingface.co/papers/2508.10894) for masked autoencoding on multimodal Earth observation data. You can find the MAESTRO model's code on its [GitHub repository](https://github.com/ignf/maestro).

A minimal example for using FLAIR-HUB with the MAESTRO framework:
```bash
poetry run python main.py \
model.model=mae \
model.model_size=medium \
run.exp_name=mae-m_flair \
run.exp_dir=/path/to/experiments/dir \
datasets.root_dir=/path/to/dataset/dir \
datasets.flair.rel_dir=FLAIR-HUB \
datasets.filter_pretrain=[flair] \
datasets.filter_finetune=[flair]
```
<hr>

## 📚 How to Cite

```
Anatol Garioud, Sébastien Giordano, Nicolas David, Nicolas Gonthier.
FLAIR-HUB: semantic segmentation and domain adaptation dataset. (2025).
DOI: https://doi.org/10.48550/arXiv.2506.07080
```

```bibtex
@article{ign2025flairhub,
doi = {10.48550/arXiv.2506.07080},
url = {https://arxiv.org/abs/2506.07080},
author = {Garioud, Anatol and Giordano, Sébastien and David, Nicolas and Gonthier, Nicolas},
title = {FLAIR-HUB : Large-scale Multimodal Dataset for Land Cover and Crop Mapping},
publisher = {arXiv},
year = {2025}
}
```

## ⚙️ Acknowledgement

Experiments have been conducted using HPC/AI resources provided by GENCI-IDRIS (Grant 2024-A0161013803, 2024-AD011014286R2 and 2025-A0181013803).